Annotating neurophysiologic data at scale with optimized human input.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Neuroscience research generates vast datasets, posing annotation challenges.
• Expert annotation is time-consuming and lacks reproducibility.
• Current automated methods require extensive labeled data, limiting scalability.

Purpose of the Study:

• To develop a semi-automated annotation approach for intracranial EEG (iEEG) data.
• To improve efficiency and scalability in iEEG annotation.
• To present Annotation Co-pilot, a human-in-the-loop (HITL) solution.

Main Methods:

• Utilized deep active learning and self-supervised learning (SwAV) for iEEG annotation.
• Trained a deep neural network on 1,500 hours of unlabeled iEEG recordings.
• Employed active learning to prioritize informative data epochs for expert review.

Main Results:

• Analyzed over 80,000 iEEG clips (1,176 hours).
• Achieved performance comparable to top seizure detectors using 1/6th the human annotations (AUC 0.9628 ± 0.015).
• Demonstrated superior consistency over human annotators (Cohen's Kappa 0.95 ± 0.04).

Conclusions:

• Annotation Co-pilot achieves expert-level performance in iEEG annotation.
• The method reduces annotation time by 83%, enhancing efficiency.
• Shows promise for accelerating electrophysiology research and clinical translation of AI.